[BugFix] fix cpu prefix cache bug (#5544)

* fix_dy_c8_bug

* add block_num check

* fix test case

* update ci case

fastdeploy/cache_manager/cache_transfer_manager.py

@@ -66,6 +66,13 @@ def parse_args():
         choices=["uint8", "bfloat16", "block_wise_fp8"],
         help="cache dtype",
     )
+    parser.add_argument(
+        "--default_dtype",
+        type=str,
+        default="bfloat16",
+        choices=["float16", "bfloat16", "uint8"],
+        help="paddle default dtype, swap_cache_batch only support float16、bfloat16 and uint8 now",
+    )
     parser.add_argument("--key_cache_shape", type=str, default="", help="key cache shape")
     parser.add_argument("--value_cache_shape", type=str, default="", help="value cache shape")
     parser.add_argument("--cache_queue_port", type=int, default=9923, help="cache queue port")
@@ -124,6 +131,7 @@ class CacheTransferManager:
         self.num_gpu_blocks = self.key_cache_shape[0]
         self.num_extra_layers = self.speculative_config.num_extra_cache_layer
         self.num_extra_layer_gpu_blocks = int(self.num_gpu_blocks * self.speculative_config.num_gpu_block_expand_ratio)
+        paddle.set_default_dtype(args.default_dtype)
 
         self.swap_to_cpu_thread_pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)
         self.swap_to_gpu_thread_pool = concurrent.futures.ThreadPoolExecutor(max_workers=1)

fastdeploy/cache_manager/prefix_cache_manager.py

@@ -279,6 +279,7 @@ class PrefixCacheManager:
                 + f" --local_data_parallel_id {self.local_data_parallel_id}"
                 + f" --rdma_port {cache_config.rdma_comm_ports[i] if cache_config.rdma_comm_ports is not None else '0'}"
                 + f" --speculative_config '{self.speculative_config.to_json_string()}'"
+                + f" --default_dtype '{self.config.model_config.dtype}'"
                 + (" --create_cache_tensor" if create_cache_tensor else "")
                 + f" >{log_dir}/launch_cache_transfer_manager_tprank{i}.log 2>&1"
             )